Monitoring of drilling for burr detection using Machine Learning techniques
Developing a monitoring control system for burr formation during aluminium, Al 7075-T6, drilling in order to eliminate or reduce a non-productive operation; deburring. Monitoring system based on internal signal from spindle torque capturing, to detect non desired burr formation. 10 variables were analysed as burr-related (5 cutting conditions and 5 torque magnitudes), applying Hill-Climbing for attribute selection they turned out to be 5 with an important influence on burr formation (2 cutting conditions and 3 torque magnitudes). A data base was built and later on analysed by Machine Learning in order to extract the necessary information to develop an algorithm able to detect whether the burr is over aeronautical limits or not. Suitable improvement techniques from Machine Learning were used pursuing an increment on the prediction level and an elimination of false negatives.
| Attachment | Size |
|---|---|
| Monitoring burr by machine learning techniques.wmv | 7.38 MB |
Thank you for your interest in our work.
During the previous experimentation we needed to see the signal clearer enough to detect a posible relationship with the burr formation, our main weapon was the signal's form not its noise. We choose a tipical filter method, like the butterworth because it was recommended and it worked, and kept its factors constant to observe all the signals with the same "eyes".
With out filtering we could not obtain similar results, our prediction capacity would be much lower.
I hope I answered your question if not, please do not hesitate to ask again.
Carles Argelich
Thank you Carles,
May I ask why you did not consider using the rules generated by PRISM as from Table 4 in your paper, it appears that the mean value obtained with that algorithm is as good as those obtained with the classification trees? I think that the model created with inductive learning algorithms are also user intuitive.
Emmanuel
Thank you very much for your interest Emmanuel,
Yes, you are right, the rules generated by PRISM had a higher prediction level and the induction rules are also intuitive, but they also create other rules not so intuitive and they make it hard to interpret and not so transparent as the classification trees, the Trees are easy to embed, understand and program and they let us see which variables are the most valuable. Besides, after the improvement techniques the final prediction level is better than the one achieved by the PRISM algorithm.
But, I also have to say that if our objective would had been achieving a good prediction level fast and reliable enough we would had chosen the PRISM algorithm. But as a research centre we focused the problem from another point of view.
Please, if something is not clear enough, let me know.
Carles Argelich
Thanks Carles, your answer is clear.
I am more familiar with using inductive learning algorithms as it is one of the research domain of the machine learning group at the MEC (Cardiff University). In particular, a series of such algorithms, called the RULES family of algorithms, have been developed in the centre over the past 10 years. For example, last year I used RULES-5 for the problem of recognising manufacturing features in CAD models (IPROMS2006 - concurrent engineering session). If you wish to investigate additional algorithms for your work I think my colleagues who developped the RULES algorithms would be happy to help you. The algorithms can deal with both discrete class values (as in your case) and also continuous classe values.
Best wishes,
Emmanuel
Dear authors,
Thank you for this interesting presentation. As you mentioned in the paper the influence of the Cutting speed and Feed rate chosen for the process is very important (or most important). Could you tell us a bit more about the criteria to choose the values shown in Table 1, especially for Cutting speed values.
Normally the ratio for Vc uncoated/Vc coated is 1/2. Did you choose a little bit smaller Vc on purpose?
Best regards,
Krastimir
Tnak you for your interest Krastimir,
The cutting conditions were not small for dry drilling of this aluminium alloy, in fact some of them were a little bit too high for this operation, specially if you combine it with a low feed rate, because it heats up too much.
The values were chosen after some previous experimentation: starting at the tool's manufacturer recommended conditions and searching the best conditions trying to achieve short times per hole.
Those are our reasons, but if you dry-drill aluminium faster, could you tell me which conditions do you use? We would really appreciate the information and deffinetly we would try it.
Best regards,
Carles Argelich
Thank you very much Emmanuel, in case we decide to investigate further in this area we will in deed contact your colleagues because it sounds interesting for a further research. In fact, new projects pointing at machine learning algorithms are coming up in our centre.
Best wishes,
Carles.
Dear Carlos (or Carles!),
The main reason for my asking was caused by the missing explanation HOW the regimes were chosen.
On the other hand this parameters are the main cause of burrs and they have to be optimised somehow, as you mentioned from the tool manufacturer data + the experimental work already done.
This could be mentioned even in the references.
Best regards,
Krastimir
Dear Carlos,
It looks like you paper is generating a fair bit of interest. I am wondering if you tried to separate the 106 data in a training set and a testing set. If yes, which percentage of data did you use to build both sets?
Thanks,
Emmanuel
Dear Emmanuel,
Sorry for the delay and thank you, once again.
The data was divided into 10 aleatory parts. On each one of these parts, 9 pieces were used for training and one for testing. So, the percentage you ask for would be 90 - 10 for training - testing.
Best regards,
Carles.
Dear authors,
Thank you for your contribution to IPROMS2007 and for providing this very interesting presentation. Could you tell us a bit more about the importance of filtering the recorded signals with the Kaiser Window method and the Butterworth filter? Could you have obtained similar results without filtering?
Emmanuel